Related Experiment Video
Updated: Aug 27, 2026
Comparison of Predictive Performance of Three Lymph Node Staging Systems in Colorectal Signet Ring Cell Carcinoma Based on Machine Learning Model
Published on: April 18, 2025
Prediction of Lymph Node Metastasis in Early Gastric Cancer Using Foundation Model Ensembles and Patch-Based
Woojin Chung1, Yujun Park2, Yoonjin Kwak3
1Department of Biomedical Engineering, Hankuk University of Foreign Studies, Yongin, Korea.
Purpose:
Accurate prediction of lymph node metastasis (LNM) remains challenging in early gastric cancer (EGC), particularly when determining the need for additional surgery after endoscopic resection. We developed and evaluated a deep learning framework using multiple pathology foundation models for LNM prediction and explored histological features associated with model-predicted metastatic risk.
Materials And Methods:
Foundation model ensembles were developed using whole-slide images from surgical EGC cases and evaluated in internal and external surgical cohorts. The model was further tested in an endoscopic submucosal dissection cohort in which all patients subsequently underwent additional gastrectomy with lymph node dissection. Performance was assessed by area under the curve (AUC) and compared with scratch-trained baseline models. At a fixed sensitivity of 1.00, the potential reduction in lymph node dissection among histologically node-negative patients was evaluated. Model-guided histological analysis was performed to interpret features associated with predicted metastatic risk.
Results:
The final ensemble model achieved an AUC of 0.879 (95% CI, 0.799-0.944) in the internal and 0.801 (95% CI, 0.697-0.891) in the external cohort, representing improvements of 0.040 and 0.066, respectively, over the baseline model. In the endoscopic submucosal dissection cohort, the model achieved an AUC of 0.766 (95% CI, 0.635-0.882), representing an improvement of 0.189 over the baseline model. At a fixed sensitivity of 1.00, the model suggested avoiding lymph node dissection in approximately 5% of histologically node-negative patients. Model-guided histological analysis showed associations consistent with established pathological features, with exploratory observations including context-dependent associations between inflammatory infiltration and tumor differentiation, differences in predicted risk between well- and moderately differentiated tumors, and lower predicted risk in poorly cohesive carcinoma.
Conclusions:
A multiple foundation model ensemble showed improved LNM risk prediction in EGC and provided interpretable histological insights supporting clinical understanding and risk stratification.
